There exists a large gap in the soft ground settlement between theory prediction and the practice，so it is difficult to meet the design requirements and is not conducive to guide the construction. Combination of existing theoretical prediction methods and field observation information is helpful to control the engineering construction. With the combination of genetic algorithm and BP optimization method to train the network, weights of BP neural network can be optimized; with Gong Paz curve to decompose the settlement timing, network was trained by offset of settlement trend line. Good application results were achieved in predicting soft ground settlement by using this method.